On-line incremental learning for vision-guided real-time navigation using improved updating

In this paper, the issue of incremental learning for autonomous vision-guided navigation by a mobile robot is investigated. The autonomous navigation task is regarded as a content-based retrieval problem where the robot learns the navigation experience using a recursive partition tree (RPT). During real-time navigation, each new image is used to retrieve the best-matched image from the tree that has been automatically constructed in the learning phase. The associated control signal of the retrieved image is used to control the next steering of the robot. The eigensubspace method is used for feature derivation and the RPT is used for fast retrieval. We introduce an incremental learning algorithm which has a real-time implementation for both learning and performance phases. Two improved versions of the incremental learning algorithms, ILA1 and ILA2, are presented to reduce the maximal response time. Experiments were conducted to verify the performance of the proposed new algorithms. The proposed incremental learning can be performed on-line in real time.